b70-optimization-lab

DeepSeek V4 K160 MTP1 row-exact record

Date: 2026-07-15

Result

The checkpoint-attached one-layer MTP is now a qualified TP4+EP record:

LocalMaxxing approved the result as cmrmetch81nm3mj01w1pidsyt.

Evidence:

/mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/mtp1-compressor-rowexact-graph-oneccl1712-20260715T1840Z

First screen was invalid

Plain MTP1 looked even faster: 50.743348 and 50.103147 tok/s across two strict suites. It passed the first seven ordered exact captures. The eighth later capture failed arithmetic-b: instead of stopping after 437, the response continued with prompt text. This was a real sustained graph-state failure, not prefix caching (cached_tokens=0) or an output-window issue.

That screen is rejected and was never submitted. Preserve it at:

/mnt/fast-ai/bench-results/deepseek-v4-flash-xpu/mtp1-graph-oneccl1712-allreduce128k-20260715T1830Z

The earlier eight-token strict-JSON miss is separately classified as harness truncation: it ended with finish_reason=length one character before the closing brace. All qualification captures use 16 output tokens.

Repair and why it works

MTP1 makes the target verifier operate at M=2. DeepSeek V4’s FP32 compressor projections normally execute one M=2 torch.mm, whose row behavior on this XPU graph path did not preserve the established M=1 arithmetic/state boundary. The default-off VLLM_XPU_V4_COMPRESSOR_M2_ROW_EXACT=1 route executes each compressor as two M=1 FP32 projections and concatenates the results. This is vLLM commit 25f5ac6ee0ab872b8e9ace8d299ab28aa3fc555a, contained in the promoted head 93fde4186.

With the flag enabled, ten exact captures passed before the strict suites and another ten passed after them. The latter is important: the uncorrected path failed only after sustained request history. The repaired result therefore has a stronger replay gate than the initial fast screen.

The model loads 25.49 GiB of weights per rank, shares the target embedding, LM head, and top-k buffer with the draft, leaves 4,658 KV tokens at the 1K setting, and captures the M=2 decode graph. All four workers map the selected oneCCL 2021.17.2 runtime with SHA-256 bf9f2e447ebb8373d7b0a41b15a32f2518a670a107d6f8fd84474798d8f2301b. The 128 KiB large-prefill-allreduce route remains active.

Decision

Promote MTP1 row-exact as the current speed record while keeping the 40.170350 tok/s nonspeculative result as the base authority. The result is target verified and is not aggregate throughput. K160 remains a hash-pruned performance checkpoint, not the final quality-selected DeepSeek V4 pack.

The next bounded ceiling test is MTP2, which reuses the one attached MTP layer twice. It proceeds only if exact post-load replay passes and complete-cycle throughput beats MTP1; lower acceptance or another M=3 state defect closes it.